Reproducible Research

History of reproducible research

• Mathematics (400 BC?)
• Write scientific paper, Galileo, Pasteur, etc. (1660s?)
• Publish a pidgin algorithm and describe simulation datasets (1950s?)
• Sell magtape of code and data (1970s?)
• Place idiosyncratic dataset & software at website (1990s?)
• Publish datasets and scripts at website, eg. biology, political science, genetics, statistics (2000s?)
• Hosted integrated code and data (2020s?)

Gavish & Gonoho AAAS 2011, Oxberry 2013

Motivations: Claerbout’s principle

“An article about computational result is advertising, not scholarship. The actual scholarship is the full software environment, code and data, that produced the result.”
Claerbout and Karrenbach, Proceedings of the 62nd Annual International Meeting of the Society of Exploration Geophysics. 1992

Benefits are straightforward

• Verification & Reliability: Find and fix bugs. Today’s results == tomorrow’s.
• Transparency: increased citation count, broader impact, improved institutional memory
• Efficiency: Reduces duplication of effort. Payoff in the (not so) long run
• Flexibility: When you don’t ‘point-and-click’ you gain many new analytic options.

But limitations are substantial

Technical

• Classified/sensitive/big data
• Nondisclosure agreements & intellectual property
• Competition
• Neither necessary nor sufficient for correctness (but useful in disputes)

Cultural & personal

• Few follow even minimal reproducibility standards.
• Few expect reproducibility
• No uniform standards
• Inertia & embarassment

Our work exists on a spectrum of reproducibility

Peng 2011, Science 334(6060) pp. 1226-1227

Goal: expose more of the research workflow

http://www.stodden.net/AMP2011/slides/pengslides.pdf

Case study: Reinhart and Rogoff controversy

• 2010: Claimed high debt-to-GDP ratios led to low GDP growth
• Threshold to growth at a debt-to-GDP ratio of >90%
• Substantial popular impact on autsterity politics

Excel coding error sliced several countries out of the data set…. The Economist

Case study: Gay Marrage

In May 2015 Science retracted a study of how canvassers can sway people’s opinions about gay marriage published just 5 months ago.

• Original survey data not made available for independent reproduction of results
• Survey incentives misrepresented
• Sponsorship statement false.

- Science Editor-in-Chief Marcia McNutt

Two Berkeley grad students who attempted to replicate the study quickly discovered that the data must have been faked.

Source

Case study: Seizure Medicine

2013 Seizure study retracted after authors realize data got “terribly mixed”

“The article has been retracted at the request of the authors. After carefully re-examining the data presented in the article, they identified that data of two different hospitals got terribly mixed. The published results cannot be reproduced in accordance with scientific and clinical correctness.”" Authors of Low Dose Lidocaine for Refractory Seizures in Preterm Neonates

Source

Bad spreadsheet merge kills depression paper, quick fix resurrects it

Authors informed the journal that the merge of lab results and other survey data used in the paper resulted in an error regarding the identification codes. Results of the analyses were based on the data set in which this error occurred.

“Lower levels of CSF IL-6 were associated with current depression and with future depression […]” Original conclusion

“Higher levels of CSF IL-6 and IL-8 were associated with current depression […]” Revised conclusion

Source

Literate statistical programming

“Instead of imagining that our main task is to instruct a computer what to do, let us concentrate rather on explaining to humans what we want the computer to do.” Donald E. Knuth, Literate Programming, 1984

For example… If I say that 2 + 2 = 4, why would you believe me?

2+2

## [1] 4

The text and R code are interwoven in the output. The products are ephemeral - focus is on the methods!

Literate statistical programming

Of course 2+2 is trivial, but what if I said:

Depositing code and data

Payoffs - Free space for hosting (and paid options) - Assignment of persistent DOIs - Tracking citation metrics

Costs - Sometimes license restrictions (CC-BY & CC0) - Limited or no private storage space

A hierarchy of reproducibility

• Good: Use code with an integrated development environment (IDE). Minimize pointing and clicking (RStudio)
• Better: Use version control. Help yourself keep track of changes, fix bugs and improve project management (RStudio & Git & GitHub or BitBucket)
• Best: Use embedded narrative and code to explicitly link code, text and data, save yourself time, save reviewers time, improve your code. (RStudio & Git & GitHub or BitBucket & rmarkdown & knitr & data repository)

Stodden (IASSIST 2010) sampled American academics registered at the Machine Learning conference NIPS (134 responses from 593 requests (23%). Red = communitarian norms, Blue = private incentives

Stodden (IASSIST 2010) sampled American academics registered at the Machine Learning conference NIPS (134 responses from 593 requests (23%). Red = communitarian norms, Blue = private incentives

Standards to normalise reproducible research

Reproducible Research Standard (Stodden 2009)

• The full compendium on the internet
• Media such as text, figures, tables with Creative Commons Attribution license (CC-BY)
• Code with one of Apache 2.0, MIT, LGPL, BSD, etc.
• Original “selection and arrangement” of data with CC0 or CC-BY

Biggest challenge: culture change

Promote culture change through positive attribution

Implement mechanisms to indicate & encourage degrees of compliance (ie. clear definitions for different levels of reproducibility), cf. Stodden’s:

• ‘Reproducible’: compendium of text-code-data online
• ‘Reproduced’: compendium available and independently reproduced
• ‘Semi-Reproducible’: when the full compendium is not released
• ‘Semi-Reproduced’: independent reproduction with other data
• ‘Perpetually Reproducible’: streaming data

A universal document converter, open source, cross-platform

• Write code and narrative in rmarkdown
  
• knitr->markdown (with computation)
  
• use pandoc to get HTML/PDF/DOCX

Tracking changes with version control

Payoffs

• Eases collaboration
• Can track changes in any file type (ideally plain text)
• Can revert file to any point in its tracked history

Costs - Learning curve

Final

Abandoning the habit of secrecy in favor of process transparency and peer review was the crucial step by which alchemy became chemistry.
Raymond, E. S., 2004, The art of UNIX programming: Addison-Wesley.

Colophon

• Slides based on Ben Marwick’s presentation to the UW Center for Statistics and Social Sciences (12 March 2014) (OrcID)

References: See Rmd file for full references and sources